SenseWhy: Overeating in Obesity Through the Lens of Passive Sensing

项目摘要/摘要 医学专业人士最近打消了这样一种观点，即每个人都有理想的减肥饮食。一 肥胖的原因是暴饮暴食，但我们不知道是什么模式和行为导致了这一问题 习惯。确定导致能量失衡的有问题的饮食行为对于治疗肥胖症至关重要。 研究通常集中在单一的假定的暴饮暴食的原因机制上，如压力或渴望，而不是 解决与暴饮暴食同时出现的多种特征。因此，预测暴饮暴食的因素 插曲仍然未知，它们中的哪些有助于个体的一致性和可变性 暴饮暴食。 鉴于最近在被动感知方面的进步，我们现在有可能通过使用 无缝捕捉生理特征，如喂食手势和燕子的数量以及心率 可变性。收集识别暴饮暴食的可检测和可预测的特征将磨练出 干预者可能最好的目标是帮助肥胖人群了解他们的饮食习惯，并最终 提高他们自我调节饮食行为的能力。区位尺度模型将映射出大多数 帮助受试者养成习惯，为干预者提供指导行为的基本目标。 第一个目标是收集基于传感器的生态瞬时评估数据(评估尚未确定的因素 可通过传感检测)，并应用机器学习算法来识别子集 检测暴饮暴食的功能，根据摄像的进食事件和24小时的基本事实进行验证 饮食回忆。参与者将佩戴被动传感传感器套件，并对随机和事件触发做出反应 关于每一集进食的提示。然后，机器学习将确定最优特征子集 使用梯度助推器检测暴饮暴食发作。在第二个目标中，层次聚类 技术将把暴饮暴食的发作归类为理论上有意义的和临床上已知的问题 与暴食有关的行为。最终的目标是建立统计模型来解释可检测到的 以及临床上已知的关于新习惯养成的问题特征。这些模型将为 优化研究以发现可指导及时干预治疗肥胖症的循证决策规则 通过防止暴饮暴食，保持健康的饮食行为。

PROJECT SUMMARY/ABSTRACT Medical professionals have recently put to rest the idea that there is an ideal weight loss diet for everyone. One cause for obesity is overeating, but we do not know what patterns and behaviors contribute to this problematic habit. Defining problematic eating behaviors that lead to energy imbalance is essential for treating obesity. Studies typically focus on a single putative causal mechanism of overeating such as stress or craving, not addressing the multiple features that co-occur with overeating. Hence, the factors that predict overeating episodes remain unknown, as do which of them contribute to an individual's consistency and variability of overeating. Given recent advancements in passive sensing, we now have the potential to detect problematic eating using seamlessly captured physiological features such as number of feeding gestures and swallows, and heart rate variability. Collecting detectable and predictable features that identify overeating will hone in on the patterns that interventionists may optimally target to help populations with obesity understand their eating habits and ultimately improve their ability to self-regulate their eating behaviors. Location-scale models will map the factors that most contribute to habit formation within subjects, providing interventionists with essential targets to guide behavior. The first aim is to collect sensor-based and ecological momentary assessment data (to assess factors not yet detectable through sensing) from adults with obesity and apply machine learning algorithms to identify a subset of features that detect overeating, as validated against ground truth of videotaped eating episodes and 24 hour dietary recall. Participants will wear a passive sensing sensor suite and respond to random and event-triggered prompts regarding each eating episode. Then, machine learning will determine the optimal feature subset that detect overeating episodes using Gradient Boosting Machines. In the second aim, hierarchical clustering techniques will cluster overeating episodes into theoretically meaningful and clinically known problematic behaviors related to overeating. The final aim is to build statistical models that explain the effect of detectable and clinically-known problematic features on new habit formation. These models will lay a foundation for optimization studies to discover evidence-based decision rules that can guide timely interventions to treat obesity by preventing overeating, and maintaining healthy eating behaviors.